This report presents EXAONE Finance, a financial time series foundation model designed for forecasting. It addresses limitations of existing TSFMs, which rely on self-attention and are computationally costly for long, multi-channel data.
EXAONE Finance replaces self-attention with a causal 1D convolution for temporal mixing and a group-aware pooling MLP for variate mixing, enabling linear-time processing. It also uses masked context augmentation during training to handle missing data common in financial markets.
The model is pretrained on a large-scale financial corpus covering equities, foreign exchange, commodities, crypto-assets, fixed income, and macroeconomic indicators. It achieves state-of-the-art results on the FinVerse benchmark across point-forecast accuracy, asset ranking, and portfolio profitability.
This approach improves efficiency and robustness for financial forecasting tasks, making it suitable for long, many-channel, intermittently observed data.
Source: https://arxiv.org/abs/2609.04239